Concurrency – a cautionary tale
Grand Central Dispatch (GCD) by Apple
API for lightweight threads
Work queues hold functions (or code blocks) to be called
Queues can manage concurrency, coordinate joining of tasks
GCD manages scheduling and processor assignment
A simple practical example:
Generating image thumbnails, for all files in a directory, on a quad-core machine.
http://www.mikeash.com/pyblog/friday-qa-2009-09-25-gcd-practicum.html
v1: Sequential program – 984 seconds
v2: Naïve parallelisation – did not run
GCD spawns another thread each time one blocks
Each task (a) hits the disk, (b) allocates memory to decompress images
Then the OS starts swapping, i.e. more disk contention
and GCD spawns another thread…
v3: Read files sequentially in a single task, compress in parallel – 300 seconds
This is self-limiting, but can still thrash
(Used only 10GB RAM on a machine with 15GB!)
v4: Limit the number of tasks to 2 x number of CPU cores – 279 seconds
The CPU is not the only resource that threads / processes must share.
Concurrency (and distribution) libraries provide programmer support, not solutions.
Cross address-space IPC
16